• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Poisoning attack detection scheme based on data integrity sampling audit algorithm in neural network

    2023-12-05 07:36:40ZhaoNingningJiangRui

    Zhao Ningning Jiang Rui

    (School of Cyber Science and Engineering,Southeast University,Nanjing 210096,China)

    Abstract:To address the issue that most existing detection and defense methods can only detect known poisoning attacks but cannot defend against other types of poisoning attacks,a poisoning attack detecting scheme with data recovery (PAD-DR) is proposed to effectively detect the poisoning attack and recover the poisoned data in a neural network.First,the PAD-DR scheme can detect all types of poisoning attacks.The data sampling detection algorithm is combined with a real-time data detection method for input layer nodes using a neural network so that the system can ensure the integrity and availability of the training data to avoid being changed or corrupted.Second,the PAD-DR scheme can recover corrupted or poisoned training data from poisoning attacks.Cauchy Reed-Solomon (CRS) code technology can encode training data and store them separately.Once the poisoning attack is detected,the original training data is recovered,and the system may get data from any k nodes from all n stores to recover the original training data.Finally,the security objectives of the PAD-DR scheme to withstand poisoning attacks,resist forgery and tampering attacks,and recover the data accurately are formally proved.

    Key words:poisoning attack; neural network; deep learning; data integrity sampling audit

    Deep learning (DL)[1]is the foundation for several modern artificial intelligence applications.It has rapidly matured and entered safety-critical applications such as self-driving cars[2-3],drones,and robots[4-5].Deep neural network (DNN)[6]is a machine learning technique that tries to imitate the neurons of the human brain to transmit information and interpret data.

    Unfortunately,the defense methods against poisoning attacks are not systematic enough.Zhang et al.[14]designed a strong defense scheme called DUTI to deal with the label-flipping attack.Given a small portion of the trusted items,the DUTI scheme could learn the difference between the distribution of trusted items and training samples to find the potentially corrupted labels and then get the corrupted labels to a domain expert for further examination.To defend against tool command language (TCL) attacks[12],Peri et al.[15]proposed a scheme named Deep-kNN to remove poisoning samples.In Ref.[15],the authors compared the class labels of each testing sample with itskneighbors.If most neighbor samples differed from the testing sample,this testing sample should be removed.Using the cooperative deep learning system,Shen et al.[16]proposed a defense method called AUROR to defend against poisoning attacks.The AUROR scheme could automatically identify and display the process of abnormal distribution features and then detect malicious users in the system according to the abnormal features.Based on a prior observation that poisoned samples may exploit spare capacity in the neural network,Liu et al.[17]proposed a fine-pruning technology as a defense method to enhance the security of deep neural networks by eliminating some dormant neurons to turn off poisoning behaviors.

    Considering that the previous schemes[14-17]can only defend against specific attacks,Diakonikolas et al.[18]proposed a robust algorithm named Server that could defend against poisoning attacks in classification and regression models.Chen et al.[19]proposed a generic and attack-agnostic defense approach called De-Pois.The key idea of De-Pois was to train a mimic model to imitate the behavior of the target model with clean samples.

    In this study,we proposed a poisoning attack-detecting scheme with data recovery (PAD-DR).First,we designed a data sampling audit algorithm and combined it with a real-time data detection method to detect all kinds of poisoning attacks.Second,we applied Cauchy Reed-Solomon (CRS) code technology to encode training data and store them on multiple servers to recover the corrupted training data.Finally,we formally proved the security goals of our PAD-DR scheme to withstand poisoning attacks and to recover the data accurately.

    1 Preliminaries

    1.1 Bilinear mapping

    Definition1LetG1andG2be multiplicative cyclic groups of a large prime orderp; a pairing is a bilinear mape:G1×G1→G2.It satisfies the following properties:

    1) Bilinear.e(ua,vb)=e(u,v)ab,?u,v∈G1;a,b∈Zp.

    2) Non-degeneracy.?u,v∈G1,thuse(u,v)≠1∈G1.

    3) Computability.?u,v∈G1; there is a polynomial time algorithm to calculatee(u,v).

    4) Safety.It is difficult to calculate the discrete logarithm problem inG1andG2.

    1.2 Discrete logarithm problem (DLP)

    1.3 Computational Diffie-Hellman problem

    1.4 Co-computational bilinear Diffie-Hellman problem (CO-CDH)

    2 Proposed PAD-DR Scheme

    2.1 System model

    In our PAD-DR scheme,as shown in Fig.1,the system consists of five types of entities: system administrator (SA),third party auditor (TPA),coded data stores (CDSi),training data store (TDS),and nodes of the neural network for input layer (NNs).SA is responsible for processing the original training data.TPA is responsible for generating challenges to detect whether poisoning attacks are launched.

    Fig.1 System model for Our PAD-DR scheme

    CDSiare cloud spaces for storing regenerating backup data,where the regenerating backup data are distributed onnmultiple CDSi.TDS stores the original training data with tags and sends them to the NNs.NNs receive the training data and feed them to the neural network.Besides,NNs are responsible for detecting the poisoning attacks in real-time with the data tags.

    2.2 Threat model

    Our PAD-DR scheme defines the threat model in terms of the SA,TPA CDSi,TDS,NNs,and network attackers (NAs).

    The SA is credible.The SA encodes the training data,generates tags for the original data,and collects regenerating data blocks to help recover the poisoned data.

    The TPA is semitrusted.The TPA may run the algorithm and protocol in the system.However,the TPA is curious about the contents of training data and tries to obtain the contents of training data in the verification process.

    The TDS is semitrusted.The TDS may faithfully run the algorithm and protocol in the system.However,to maintain the reputation,they may conceal the truth when poisoning samples attack the training data.At that time,TDS may attempt to forge proofs to cheat TPA for passing the auditing process.CDSiare credible and may store encoded regenerating data for backup.NNs are credible and may run the algorithm and protocol in the system and transmit the training data to the nodes of the next layer.

    NAs are malicious.NAs attempt to launch poisoning sample attacks.Furthermore,to pass the audit by TPA and NNs,NAs may attempt to launch tempering and forgery attacks by forging data and signature proofs.

    3 Construction of PAD-DR Scheme

    We proposed the detailed construction of our PAD-DR scheme.The scheme includes three phases: the setup,detection and recovery phases.Details of each phase are as follows.

    3.1 Setup phase

    The setup phase includes setup,encoding and SigGen algorithms.Among them,the setup algorithm generates system parameters,the encoding algorithm applies CRS code technology to encode the original training data for backup,and the SigGen algorithm generates tags for the original training data.Three algorithms are described in detail as follows.

    3.1.1 Setup algorithm

    3.1.2 Encoding algorithm

    Supposem,k,w∈Z+,a Galois field GF(2w),for data fileF={m1,m2,…,mk},SA utilizes CRS[22]technology to generaten=m+kencoded data blocks by constructing a code matrixC=ΨM,where the encoding matrixΨisn×k,and the message matrixMisk×1.

    Then,the encoding matrixΨis designed as follows:

    Hence,thei-th row ofΨis defined as the encoding vectorΨi(i∈{1,2,…,n}).

    For example,letk=6,m=4,andn=10,the Cauchy matrixGis with dimensions 4×6 on GF(28).LetX={0,1,2,3,4,5},Y={6,7,8,9},the encoding process is shown as follows:

    C=Encode(M)=ΨM=

    3.1.3 SigGen

    3.2 Detection phase

    The detection phase consists of TPA detection and node detection algorithms.With our proposed sampling audit algorithm,the TPA detection algorithm can detect poisoning attacks on the training data stored in TDS.In real-time,a node detection algorithm can detect poisoning attacks on the training data at the input layer for NNs.

    3.2.1 TPA detection

    Having received the response proof from the TDS,the TPA verifies the proof as follows.TPA checks the verification equation as

    (1)

    If Eq.(1) holds,the training data stored on TDS are securely protected.Otherwise,the training data stored on TDS should be changed or corrupted by the poisoning attacks.When the training data is detected to be poisoned,the TPA immediately makes an alarm for the poisoning attack and sends feedback {error} to the SA.Then,the SA performs the recovery operation to recover the poisoned data on TDS.

    3.2.2 Node detection

    After receiving training data with signatures from TDS,NNs run a node detection algorithm to detect poisoning attacks as follows.First,NNs verify the correctness of the signatureStsk(F‖μ′‖σ′) by TDS’s public key according to the RSA signature algorithm.If the verification fails,the NNs abort the message and send {error} to SA.Otherwise,we have to design the real-time data detection method for the NNs to check the verification equation as follows:

    (2)

    If Eq.(2) holds,the training data sent by TDS are complete and correct.Otherwise,the poisoning attacks could change or corrupt the training data.NNs may send error feedback to inform SA that the training data could be attacked and request SA to recover the training data.

    3.3 Recovery phase

    In the recovery phase,there is a data recovery algorithm.Once SA receives error feedback from TPA or NNs,implying that the training data has been poisoned and attacked,SA may execute the data recovery algorithm to recover the original training data.

    With the data recovery algorithm,the original data fileF={m1,m2,…,mk} can be recovered by collecting anyknodes of CDSifor itsciandΨi.Thus,the original data fileFcan be recovered through linear operations on the entries of anykrows of the code matrixC.

    Next,based on the definition in the preliminaries,we provided a theorem indicating that the entire process of the PAD-DR scheme is correct.

    Theorem1Our PAD-DR scheme can correctly run to detect the poisoning attack and recover the original training data accurately.

    4 Security Analysis

    In this section,we formally proved that our PAD-DR scheme can resist tampering and forgery attacks.Also,we proved that the poisoned data can be correctly recovered.

    4.1 Related theorems

    Some theorems formally showed that our PAD-DR scheme can resist tampering and forgery attacks.Furthermore,we showed that the poisoned data can be correctly recovered.

    Theorem2In our PAD-DR scheme,it is computationally infeasible for the TDS and NAs to forge proof for passing the TPA’s auditing.Suppose there is a (t,?)-algorithm to forge a proof.Then,if a (t′,?′) adversary can solve the Co-CDH problem witht′≤t+qcG1and ?′=?/e(1+qs),whereqcG1denotes one exponentiation time inG1,andqsdenotes the number of requests.

    Theorem3In our PAD-DR scheme,NNs can detect poisoned samples in real-time training data received from TDS.

    Theorem4In our PAD-DR scheme,the original data fileFcan be recovered by collectingciandΨifrom anyknodes of CDSi.

    4.2 Security goals comparison

    In this section,the security goals of our PAD-DR scheme are compared with that of DUTI[14],Deep-kNN[15],Server[18]and De-Pois[19].The comparison includes specific poisoning attack detection,arbitrary poisoning attack detection,and poisoned data recovery.In Tab.1,“√” depicts that the security problem has been solved,“×” indicates that the problem has not been solved.

    Tab.1 Security goals comparison

    According to Tab.1,our PAD-DR scheme can realize all the security goals mentioned above,while other schemes cannot.

    5 Experiment and Performance Analysis

    This section begins by describing the experimental setup and the accuracy under different attacks.Then,we analyzed the communication overhead and computation costs in detecting poisoning attacks of training data and evaluated the performance of our PAD-DR scheme.

    5.1 Experiment setup

    5.1.1 Datasets

    The training data used in our scheme are from MNIST,CIFAR-10,and house pricing.MNIST is a database of handwritten digits containing 60 000 training sample sets and 10 000 test sample sets stored in binary form.The width and height of each sample image are 28×28.The CIFAR-10 database contains 32×32 color images in 10 different classes,with 50 000 training and 10 000 testing images.The 10 different classes include cars,dogs,frogs,airplanes,cars,ships,horses,birds,and trucks.The house pricing dataset utilizes predictor variables such as the number of bedrooms and lot square footage.It contains 1 460 houses and 81 features.In the experiment,we randomly split this dataset into training and testing datasets with 70% and 30% of the data,respectively.

    5.1.2 Experimental environment and parameter setting

    We evaluated the process of detecting poisoning samples on NVIDIA 2080Ti GPU and applied Ali Cloud to store training data.The operating system is Windows 10.The RAM size is 8 GB.Matrix multiplication algorithms are implemented with the open-source library Jerasure Version (1.2).Auditing algorithms are implemented onCwith a pairing-based cryptography library.The curve utilized in the experiment is an MNT curve.We set the length ofG1to 175.In our experiment,we adopted a common neural network that applies three full connection layers with ReLU activation and one full connection layer with sigmoid activation,and we applied the cross-entropy loss function to calculate its loss.

    5.1.3 Generation of poisoning training data

    To verify the performance of our PAD-DR scheme,we randomly extracted 10%,15%,20%,25%,and 30% of the training data to generate poisoning samples.To compare the detection rate of poisoned samples for our PAD-DR scheme,TCL[12],LF[8],and R[13]attacks are used to generate poisoning data.For MNIST,the number of iterations is 400,the learning rate is set to 0.1,and the initial value is set to 0.01.For CIFAR-10,the number of iterations is 6 000,the learning rate is 0.01,and the initial value is 0.01.

    5.2 Detection rate for poisoning attacks

    In this section,we evaluated the detection rate of poisoned samples for our PAD-DR scheme compared with that of TRIM[13],DUTI[14],Deep-kNN[15],Server[18],and De-Pois[19].The results are presented in Figs.2,3,and 4,respectively.

    As shown in Fig.2,for TCL attacks,the detection rate of our PAD-DR scheme is always 100%,which is higher than that of Refs.[15,19].There are two reasons for the relatively low detection rate of Ref.[15].One is that the authors only compared the samples with the surroundingksamples,which was not sufficient or accurate.The other reason is that the authors should set an artificial setting threshold to distinguish poisoned samples from clean samples.The main reason for the relatively low detection rate of Ref.[19]is that the authors adopted conditional GAN technology to expand a small part of a trusted dataset as the whole training data,which could not fully reflect the features of the real training data.In our PAD-DR scheme,we could only detect any changed or poisoned samples in the training data by verifying Eqs.(1) and (2).Hence,the detection rate of our PAD-DR scheme is always 100%.

    Fig.2 Detection rate under TCL attacks

    As illustrated in Fig.3,for TF attacks,the detection rate of our PAD-DR scheme is always 100%,which is higher than that of Refs.[14,18-19].In Ref.[14],the main reason for the relatively low detection rate is that the authors required a small piece of completely reliable data,which could not be ensured in the system to detect the TF attack.In Ref.[18],the detection rate decreased rapidly with the increase of the poisoning rate.The main reason to detect the TF attack is that the authors should set an artificial setting threshold,which could not accurately distinguish poisoned samples from clean samples.In Ref.[19],the main reason for the relatively low detection rate is that the authors should train the mimic model,which was ineffective in obtaining accurate prediction results.

    Fig.3 Detection rate under LF attacks

    As depicted in Fig.4,for R attacks,the detection rate of our PAD-DR scheme is always 100%,which is higher than that of Refs.[13-14,19].With the increase of the poisoning rate,the detection rate of Refs.[13-14,19]decreased rapidly.In Ref.[13],to detect R attacks,the authors ignored the influence of poisoned samples in the lowest residual set with iteration,which may cause the failure of R-attack detection.In Ref.[14],the authors required a small piece of completely reliable data to detect an R attack,which could not be ensured in the system.In Ref.[19],the authors should train the mimic model to detect R attacks,which was not effective enough to obtain accurate prediction results.In our PAD-DR scheme,we could only detect any changed or poisoned samples in the training data using verifying Eqs.(1) and (2).Hence,the detection rate of our PAD-DR scheme for R attacks is always 100%.

    Fig.4 Detection rate under R attacks

    5.3 Computation cost

    We evaluated the computation cost for our PAD-DR scheme,whereLdenotes the length of each encoded data block,EGdenotes one exponentiation inG1andG2,EZdenotes one exponentiation inZp,MGindicates a multiplication operation on groupsG1andG2,MZindicates a multiplication operation on the number fieldZp,MFrepresents a matrix multiplication in a finite fieldF(2n),MIrepresents the cost of computing the inverse of the matrixM,AZindicates an addition operation on the number fieldZp,Pedenotes the computation cost of one pairing operatione,andHdenotes the computation cost of an operation of calculating the hash value for a number.Furthermore,|I|represents the number of challenged data blocks.All the statistical results are the averages of 20 trials.

    Tab.2 depicts the computation cost of our PAD-DR scheme in the setup,detection,and recovery phases.

    Tab.2 Computation cost of our PAD-DR scheme

    In the detection phase,the computation cost refers to two parts: TPA detection and node detection.The computation cost of TPA detection is|I|(MZ+H+EG)+(|I|-1)(AZ+MG)+Pe,where|I|(MZ+H)+(|I|-1)AZand|I|EG+(|I|-1)MGare the computation cost of TDS to generate a linear combination of data blocks and an aggregated tag,respectively,wherePeis the computation cost of TPA to verify the proof.The computation cost of node detection is|I|(MZ+H+EG)+(|I|-1)(AZ+MG)+Pe,where|I|(MZ+H)+(|I|-1)AZand|I|EG+(|I|-1)MGare the computation cost of TDS to generate a linear combination of data blocks and an aggregated tag,respectively,Peis the computation cost of NNs to verify the proof.Hence,the computation cost in the detection phase is 2|I|(MZ+H+EG)+2(|I|-1)(AZ+MG)+2P.

    5.4 Communication overhead

    We assessed the communication overhead in the setup,detection,and recovery phases.In this section,|Zp|represents the size ofZp,|G|represents the size of groupG,and|GF|represents the size ofGF(2w).Tab.3 presents the communication overhead for the three phases.

    Tab.3 Communication overhead for our PAD-DR scheme

    In the setup phase,communication overhead mainly includes two parts.The one is SA uploads data and tags {F,Φ} to TDS.The other is SA sends the encoded data {ci}i=1,2,…,nto CDSi.Hence,the communication overhead in the setup phase is sizeof(F)+k|G|+n|GF|.

    In the detection phase,the communication overhead mainly refers to two parts: TPA detection and node detection.The communication overhead for TPA detection includes two parts.One is that TPA sends a challenge chal={(i,vi)} to TDS.The other is that TDS replies as proofP={μ,σ}.Hence,the communication overhead for TPA detection is 2|I||Zp|and|G|+|Zp|.The communication overhead for node detection is sizeof(F)+|G|+(k+1)|Zp|which arises from TDS sending {F,μ′,σ′,Sssk(F‖μ′‖σ′),vi}i∈Ito NNs.Thus,the total communication overhead in the detection phase is sizeof(F)+2|G|+(k+2+2|I|)|Zp|.

    In the recovery phase,the communication overhead is decided by the data size of {ci,Ψi},which arises from theknodes of CDSi.Hence,the communication overhead of our PAD-DR scheme is 2k|GF|.

    6 Conclusions

    1) An algorithm combining data sampling audit and real-time data detection is designed,and experiments show that the algorithm can accurately detect toxic data contained in the data.

    2) A data recovery algorithm based on CRS encoding is designed,and experiments show that it can efficiently restore poisoned data to clean data.

    3) The current algorithm cannot guarantee the security of parameters during transmission.We will conduct further research on improving the integrity and confidentiality of parameters in the future.

    国产免费视频播放在线视频| 男女之事视频高清在线观看 | 久久精品亚洲av国产电影网| 免费少妇av软件| 精品少妇一区二区三区视频日本电影| 最新在线观看一区二区三区 | 曰老女人黄片| 亚洲人成电影观看| 亚洲国产欧美日韩在线播放| 国产高清国产精品国产三级| 久久女婷五月综合色啪小说| 久久久久久久精品精品| 2018国产大陆天天弄谢| 人人妻人人爽人人添夜夜欢视频| 韩国高清视频一区二区三区| 亚洲伊人色综图| 满18在线观看网站| 国产日韩欧美视频二区| av电影中文网址| 精品高清国产在线一区| 自拍欧美九色日韩亚洲蝌蚪91| 欧美人与性动交α欧美精品济南到| 欧美精品人与动牲交sv欧美| 久久久久久人人人人人| 黑人猛操日本美女一级片| 日本色播在线视频| 黑丝袜美女国产一区| 一区二区日韩欧美中文字幕| 女性被躁到高潮视频| 人人妻人人澡人人看| 新久久久久国产一级毛片| 爱豆传媒免费全集在线观看| 中文字幕制服av| 亚洲成国产人片在线观看| 亚洲国产毛片av蜜桃av| 99国产精品99久久久久| 亚洲天堂av无毛| 丰满饥渴人妻一区二区三| 制服诱惑二区| 亚洲欧美色中文字幕在线| 黄片播放在线免费| 日本午夜av视频| 大型av网站在线播放| 精品国产一区二区三区四区第35| 一级毛片女人18水好多 | 9色porny在线观看| 免费看不卡的av| 热99久久久久精品小说推荐| 亚洲 国产 在线| 精品久久久精品久久久| 成人影院久久| 欧美人与性动交α欧美精品济南到| 老司机深夜福利视频在线观看 | 日本a在线网址| 宅男免费午夜| 精品人妻一区二区三区麻豆| 精品福利观看| 午夜激情av网站| 又粗又硬又长又爽又黄的视频| 欧美亚洲 丝袜 人妻 在线| 日韩精品免费视频一区二区三区| 亚洲精品国产av蜜桃| 欧美中文综合在线视频| 亚洲美女黄色视频免费看| 日韩中文字幕欧美一区二区 | 亚洲一码二码三码区别大吗| 成人午夜精彩视频在线观看| 日本欧美视频一区| 亚洲欧美色中文字幕在线| 欧美日韩视频精品一区| 考比视频在线观看| h视频一区二区三区| 亚洲,欧美精品.| 精品视频人人做人人爽| 婷婷成人精品国产| 亚洲欧美中文字幕日韩二区| 黄色a级毛片大全视频| av网站在线播放免费| 日本午夜av视频| cao死你这个sao货| 91精品三级在线观看| 男女午夜视频在线观看| 女人被躁到高潮嗷嗷叫费观| 性色av一级| 美女国产高潮福利片在线看| 国产1区2区3区精品| 丰满迷人的少妇在线观看| 青春草亚洲视频在线观看| 日韩av免费高清视频| 免费不卡黄色视频| 亚洲av日韩精品久久久久久密 | 精品福利观看| 精品久久蜜臀av无| 免费观看人在逋| 亚洲欧美色中文字幕在线| 亚洲国产欧美日韩在线播放| 久久久久精品国产欧美久久久 | 纯流量卡能插随身wifi吗| 一级黄色大片毛片| 最近手机中文字幕大全| 国产在视频线精品| 黄片播放在线免费| 精品欧美一区二区三区在线| 在线观看免费日韩欧美大片| 一级毛片黄色毛片免费观看视频| 女人高潮潮喷娇喘18禁视频| 午夜免费成人在线视频| 国产片特级美女逼逼视频| 国产成人欧美| 久久精品国产综合久久久| 午夜久久久在线观看| 亚洲欧洲精品一区二区精品久久久| 国产黄频视频在线观看| 国产男女超爽视频在线观看| 老汉色∧v一级毛片| 亚洲成人手机| 一级毛片电影观看| 各种免费的搞黄视频| 两人在一起打扑克的视频| 久久久久视频综合| 欧美大码av| 好男人视频免费观看在线| 99国产精品99久久久久| 我要看黄色一级片免费的| 欧美精品啪啪一区二区三区 | 成年美女黄网站色视频大全免费| 成年人免费黄色播放视频| 亚洲成人免费av在线播放| 十八禁高潮呻吟视频| 中文字幕精品免费在线观看视频| 操美女的视频在线观看| 国产精品偷伦视频观看了| 天天操日日干夜夜撸| 韩国精品一区二区三区| 欧美精品av麻豆av| 日本wwww免费看| 国产国语露脸激情在线看| 精品国产一区二区久久| 少妇人妻久久综合中文| 免费久久久久久久精品成人欧美视频| 亚洲欧洲日产国产| 欧美日韩视频高清一区二区三区二| 亚洲精品av麻豆狂野| 色播在线永久视频| 亚洲情色 制服丝袜| 亚洲精品国产色婷婷电影| 成年女人毛片免费观看观看9 | 亚洲精品中文字幕在线视频| 狂野欧美激情性bbbbbb| 少妇 在线观看| 亚洲国产精品成人久久小说| 精品少妇久久久久久888优播| 多毛熟女@视频| 免费久久久久久久精品成人欧美视频| 国产激情久久老熟女| 一区福利在线观看| 好男人视频免费观看在线| 日韩av免费高清视频| 制服人妻中文乱码| 国产精品一区二区在线不卡| 欧美日韩福利视频一区二区| 满18在线观看网站| 久久人妻福利社区极品人妻图片 | 国产精品九九99| 自线自在国产av| 97在线人人人人妻| 亚洲精品一二三| 亚洲精品美女久久久久99蜜臀 | 日本猛色少妇xxxxx猛交久久| 黄片播放在线免费| 亚洲欧美一区二区三区久久| 热99国产精品久久久久久7| 欧美亚洲日本最大视频资源| 久久人人爽人人片av| 性色av一级| 久久久久久久大尺度免费视频| 精品亚洲乱码少妇综合久久| 国产爽快片一区二区三区| 亚洲精品国产区一区二| 日韩一区二区三区影片| 性色av一级| 欧美亚洲日本最大视频资源| 晚上一个人看的免费电影| 国产亚洲精品第一综合不卡| 亚洲av成人不卡在线观看播放网 | 亚洲精品av麻豆狂野| 午夜91福利影院| 午夜免费鲁丝| 欧美人与性动交α欧美软件| 久久国产精品影院| 久久久久久久国产电影| 精品人妻1区二区| 中文字幕亚洲精品专区| 热re99久久精品国产66热6| 新久久久久国产一级毛片| 欧美97在线视频| 国产野战对白在线观看| 日韩 欧美 亚洲 中文字幕| 国产xxxxx性猛交| 国产精品香港三级国产av潘金莲 | 性色av乱码一区二区三区2| 国产欧美日韩一区二区三 | 青春草亚洲视频在线观看| 2018国产大陆天天弄谢| 中国国产av一级| 99国产精品一区二区三区| 三上悠亚av全集在线观看| 国产野战对白在线观看| 另类亚洲欧美激情| 日韩,欧美,国产一区二区三区| 黄频高清免费视频| 超色免费av| 一本色道久久久久久精品综合| 操出白浆在线播放| 欧美日韩一级在线毛片| 亚洲伊人久久精品综合| 免费不卡黄色视频| 十分钟在线观看高清视频www| 性少妇av在线| 在线av久久热| 日韩大片免费观看网站| av网站在线播放免费| 午夜激情久久久久久久| 美女扒开内裤让男人捅视频| 一本久久精品| 老司机亚洲免费影院| 美国免费a级毛片| 久久久久网色| 一个人免费看片子| 国产成人精品久久久久久| 国产成人一区二区三区免费视频网站 | 在线观看www视频免费| 欧美成人精品欧美一级黄| 亚洲av在线观看美女高潮| 人人妻人人爽人人添夜夜欢视频| 一区福利在线观看| a级片在线免费高清观看视频| videosex国产| 欧美精品一区二区大全| 99精品久久久久人妻精品| 亚洲五月婷婷丁香| 嫁个100分男人电影在线观看 | xxxhd国产人妻xxx| 亚洲中文日韩欧美视频| 9191精品国产免费久久| 亚洲自偷自拍图片 自拍| 亚洲国产欧美网| 又黄又粗又硬又大视频| 免费看av在线观看网站| 无遮挡黄片免费观看| 中文字幕人妻丝袜一区二区| 伊人久久大香线蕉亚洲五| 欧美97在线视频| 人妻一区二区av| 欧美在线一区亚洲| 可以免费在线观看a视频的电影网站| 亚洲中文日韩欧美视频| 亚洲av综合色区一区| 久久久欧美国产精品| 天天躁夜夜躁狠狠久久av| 精品国产乱码久久久久久小说| 大码成人一级视频| 国产99久久九九免费精品| 夜夜骑夜夜射夜夜干| 亚洲专区国产一区二区| 亚洲国产成人一精品久久久| 国产亚洲精品第一综合不卡| 亚洲国产精品999| 中文字幕另类日韩欧美亚洲嫩草| 又大又爽又粗| 亚洲综合色网址| 亚洲国产最新在线播放| 欧美 亚洲 国产 日韩一| 一级黄色大片毛片| 亚洲精品国产区一区二| 999精品在线视频| 精品国产一区二区三区四区第35| 久久久久精品人妻al黑| 欧美在线一区亚洲| 亚洲av成人精品一二三区| 久久热在线av| 高清欧美精品videossex| 成年av动漫网址| 欧美黑人欧美精品刺激| 成在线人永久免费视频| 大香蕉久久成人网| 国产精品三级大全| 女性被躁到高潮视频| 乱人伦中国视频| 另类亚洲欧美激情| 国产精品av久久久久免费| 美女大奶头黄色视频| 一区二区三区精品91| 七月丁香在线播放| 成人影院久久| 亚洲精品国产av成人精品| 晚上一个人看的免费电影| 男人添女人高潮全过程视频| 成在线人永久免费视频| 久久久久久久久久久久大奶| 亚洲熟女精品中文字幕| 国产精品秋霞免费鲁丝片| 91麻豆av在线| 国产99久久九九免费精品| 无限看片的www在线观看| 国产欧美日韩一区二区三区在线| 国产深夜福利视频在线观看| 久久久久久免费高清国产稀缺| 国产亚洲av片在线观看秒播厂| 久久人人爽人人片av| 国产精品99久久99久久久不卡| 亚洲午夜精品一区,二区,三区| 91精品伊人久久大香线蕉| 少妇粗大呻吟视频| 人妻 亚洲 视频| 久久久亚洲精品成人影院| 一区在线观看完整版| 欧美日韩视频精品一区| 国产成人影院久久av| 菩萨蛮人人尽说江南好唐韦庄| 国产精品一区二区免费欧美 | 国产激情久久老熟女| 欧美国产精品va在线观看不卡| 精品人妻1区二区| 国产精品一区二区在线观看99| 精品国产乱码久久久久久小说| 精品少妇久久久久久888优播| 精品少妇一区二区三区视频日本电影| 中文字幕制服av| 国产欧美日韩一区二区三 | 一二三四在线观看免费中文在| 女人高潮潮喷娇喘18禁视频| 久久亚洲国产成人精品v| 美女大奶头黄色视频| 高清av免费在线| 日韩 亚洲 欧美在线| 啦啦啦啦在线视频资源| 日韩制服丝袜自拍偷拍| 新久久久久国产一级毛片| 欧美人与善性xxx| 中文字幕另类日韩欧美亚洲嫩草| 男的添女的下面高潮视频| 久久久精品国产亚洲av高清涩受| 免费在线观看日本一区| 狂野欧美激情性bbbbbb| 三上悠亚av全集在线观看| 天天影视国产精品| 国产精品.久久久| 成年人黄色毛片网站| 久久久精品国产亚洲av高清涩受| 黑人猛操日本美女一级片| 美女国产高潮福利片在线看| 精品亚洲乱码少妇综合久久| 精品国产一区二区三区久久久樱花| 日本一区二区免费在线视频| 亚洲情色 制服丝袜| 秋霞在线观看毛片| 亚洲成人国产一区在线观看 | 国产精品三级大全| 黄色视频在线播放观看不卡| 欧美黑人欧美精品刺激| 天天躁夜夜躁狠狠久久av| 久久久国产一区二区| a级毛片黄视频| 国产伦理片在线播放av一区| 欧美在线一区亚洲| 国产男人的电影天堂91| 王馨瑶露胸无遮挡在线观看| 三上悠亚av全集在线观看| 午夜福利,免费看| 母亲3免费完整高清在线观看| e午夜精品久久久久久久| 国产成人a∨麻豆精品| 亚洲专区中文字幕在线| 国产亚洲欧美在线一区二区| 人人妻人人澡人人看| 亚洲色图综合在线观看| 亚洲 国产 在线| 午夜福利在线免费观看网站| 国产精品亚洲av一区麻豆| 中文字幕精品免费在线观看视频| 美女中出高潮动态图| 男女边摸边吃奶| 9色porny在线观看| 亚洲五月色婷婷综合| 一区福利在线观看| 一级,二级,三级黄色视频| av天堂久久9| 日韩中文字幕欧美一区二区 | 久久这里只有精品19| 亚洲国产欧美网| 国产日韩欧美亚洲二区| 精品免费久久久久久久清纯 | 久久天躁狠狠躁夜夜2o2o | 女人被躁到高潮嗷嗷叫费观| 国产xxxxx性猛交| 电影成人av| 男女国产视频网站| 啦啦啦啦在线视频资源| 女性生殖器流出的白浆| 中国美女看黄片| 免费av中文字幕在线| 久久人人爽人人片av| 国产成人一区二区在线| 成人黄色视频免费在线看| 啦啦啦在线免费观看视频4| 国产欧美日韩一区二区三区在线| 亚洲精品一二三| 免费高清在线观看视频在线观看| 可以免费在线观看a视频的电影网站| 婷婷色综合www| 午夜免费观看性视频| 免费观看a级毛片全部| 亚洲国产精品一区三区| 久久人人97超碰香蕉20202| 欧美xxⅹ黑人| 亚洲av电影在线观看一区二区三区| 国产亚洲av高清不卡| 亚洲,欧美精品.| 成人黄色视频免费在线看| 黄色毛片三级朝国网站| 日本五十路高清| 黄色视频不卡| 欧美xxⅹ黑人| 成人18禁高潮啪啪吃奶动态图| 婷婷色av中文字幕| 亚洲成av片中文字幕在线观看| 51午夜福利影视在线观看| 一级片'在线观看视频| 老司机亚洲免费影院| a级片在线免费高清观看视频| 久久精品国产亚洲av高清一级| av又黄又爽大尺度在线免费看| 激情视频va一区二区三区| 国产主播在线观看一区二区 | 亚洲欧美清纯卡通| 国产xxxxx性猛交| 国产黄频视频在线观看| av国产久精品久网站免费入址| 国产精品香港三级国产av潘金莲 | 无限看片的www在线观看| 欧美av亚洲av综合av国产av| 99re6热这里在线精品视频| 国产一区二区在线观看av| 亚洲精品日本国产第一区| 亚洲av美国av| av片东京热男人的天堂| 精品福利观看| 国产一区二区激情短视频 | 青草久久国产| 人人澡人人妻人| 亚洲伊人色综图| 亚洲国产欧美一区二区综合| 国产在视频线精品| 国产在线视频一区二区| 自拍欧美九色日韩亚洲蝌蚪91| 悠悠久久av| 亚洲一码二码三码区别大吗| 中文字幕最新亚洲高清| 啦啦啦啦在线视频资源| 波野结衣二区三区在线| 精品第一国产精品| 精品亚洲乱码少妇综合久久| 丝袜美腿诱惑在线| 国产视频首页在线观看| 在线观看人妻少妇| 如日韩欧美国产精品一区二区三区| 9191精品国产免费久久| 欧美激情 高清一区二区三区| 午夜激情久久久久久久| 极品少妇高潮喷水抽搐| 汤姆久久久久久久影院中文字幕| 黑人欧美特级aaaaaa片| 一个人免费看片子| 成人黄色视频免费在线看| 色婷婷久久久亚洲欧美| 亚洲精品美女久久久久99蜜臀 | 亚洲欧美日韩另类电影网站| 宅男免费午夜| 爱豆传媒免费全集在线观看| 欧美日韩av久久| 国产在线免费精品| 婷婷丁香在线五月| 久久精品国产a三级三级三级| 婷婷色麻豆天堂久久| 国产1区2区3区精品| 亚洲人成网站在线观看播放| 99热国产这里只有精品6| 热99国产精品久久久久久7| 日韩精品免费视频一区二区三区| 婷婷色麻豆天堂久久| 日韩免费高清中文字幕av| 亚洲,欧美,日韩| 色综合欧美亚洲国产小说| 久久人人97超碰香蕉20202| 亚洲欧美精品自产自拍| 久久午夜综合久久蜜桃| 国产又爽黄色视频| 日本色播在线视频| 亚洲国产精品成人久久小说| 日本黄色日本黄色录像| 女人精品久久久久毛片| 欧美亚洲日本最大视频资源| 国产熟女欧美一区二区| 久久久久久久久免费视频了| 欧美精品高潮呻吟av久久| 99精品久久久久人妻精品| 国产精品一区二区精品视频观看| 尾随美女入室| 99精品久久久久人妻精品| 黄色视频在线播放观看不卡| 欧美日韩亚洲综合一区二区三区_| 国产一区二区 视频在线| 波多野结衣av一区二区av| 精品熟女少妇八av免费久了| 一二三四在线观看免费中文在| 最近中文字幕2019免费版| 91老司机精品| 久久亚洲精品不卡| 免费在线观看黄色视频的| 一级黄片播放器| 99精国产麻豆久久婷婷| 免费av中文字幕在线| 男女午夜视频在线观看| 午夜av观看不卡| 手机成人av网站| 国产有黄有色有爽视频| 精品少妇久久久久久888优播| 成年动漫av网址| e午夜精品久久久久久久| 少妇粗大呻吟视频| 欧美日韩成人在线一区二区| 亚洲精品成人av观看孕妇| 丝袜美腿诱惑在线| 黄色a级毛片大全视频| 999精品在线视频| 十八禁人妻一区二区| 欧美精品亚洲一区二区| 国产成人精品久久久久久| 99久久99久久久精品蜜桃| 十八禁网站网址无遮挡| 婷婷丁香在线五月| 午夜福利视频在线观看免费| 精品熟女少妇八av免费久了| 黄网站色视频无遮挡免费观看| 国产福利在线免费观看视频| 国语对白做爰xxxⅹ性视频网站| 日本黄色日本黄色录像| 国产视频首页在线观看| 日韩 欧美 亚洲 中文字幕| 色婷婷av一区二区三区视频| 中文字幕亚洲精品专区| 观看av在线不卡| 另类精品久久| 又大又爽又粗| 又紧又爽又黄一区二区| 国产亚洲精品久久久久5区| 国精品久久久久久国模美| 国产一卡二卡三卡精品| 中文精品一卡2卡3卡4更新| 精品亚洲乱码少妇综合久久| 香蕉国产在线看| 99热全是精品| 麻豆国产av国片精品| 成年av动漫网址| 国产一区二区在线观看av| 午夜激情av网站| 亚洲视频免费观看视频| 日韩大片免费观看网站| 老司机深夜福利视频在线观看 | 晚上一个人看的免费电影| 亚洲人成电影观看| 国产精品免费视频内射| videosex国产| 99国产精品99久久久久| 欧美日本中文国产一区发布| 中文字幕最新亚洲高清| 午夜福利一区二区在线看| 久久久精品94久久精品| 欧美日韩亚洲高清精品| 国产一区二区三区av在线| 亚洲精品一卡2卡三卡4卡5卡 | 国产在线一区二区三区精| 18禁黄网站禁片午夜丰满| 日韩制服丝袜自拍偷拍| 国产一区二区在线观看av| 精品人妻1区二区| 成人国产av品久久久| 午夜福利在线免费观看网站| 国产精品一区二区在线不卡| 午夜免费观看性视频| 啦啦啦在线观看免费高清www| 国产亚洲精品第一综合不卡| 操美女的视频在线观看| 黑人巨大精品欧美一区二区蜜桃| 欧美人与性动交α欧美精品济南到| 精品国产超薄肉色丝袜足j| 青草久久国产| 国产精品久久久久久精品古装| 制服诱惑二区| 国产一级毛片在线| 69精品国产乱码久久久| 晚上一个人看的免费电影| 日本色播在线视频| 免费在线观看影片大全网站 | 亚洲欧洲国产日韩| 欧美成人精品欧美一级黄| 日本色播在线视频| 久久久久久人人人人人| 99香蕉大伊视频| 久久久久久久国产电影| bbb黄色大片| 亚洲精品一区蜜桃| 可以免费在线观看a视频的电影网站| 少妇精品久久久久久久|